This paper introduces SUPERB, a comprehensive benchmark for evaluating speech foundation models. The goal is to assess the generalizability of these models across diverse speech processing tasks. The benchmark includes 15 tasks, covering content, speaker characteristics, prosody, semantics, and generation. A unified framework is proposed, using a frozen foundation model with task-specific, lightweight prediction heads. The framework employs a learnable weighted-sum approach to combine representations from all layers of the foundation model, enabling effective performance across multiple tasks. The benchmark emphasizes the direct usability of SSL models on real-world applications, providing a standardized task design and evaluation protocol. The results show that SSL techniques are promising for building speech foundation models, with the best-performing models achieving competitive generalizability across most tasks. The benchmark also includes an online leaderboard for submissions and a community-driven database to support new development cycles. The paper analyzes the effectiveness of the framework, including information flows across tasks, the correctness of the weighted-sum benchmarking protocol, and the statistical significance and robustness of the benchmark. It also explores the performance of SSL models on different tasks, highlighting the importance of layer-wise analysis and the potential for improving performance through layer-wise benchmarking. The results demonstrate that SSL models can achieve strong task generalizability, with some models outperforming traditional non-SSL approaches. The benchmark provides a standardized evaluation framework for speech foundation models, enabling researchers to compare and improve their models effectively.This paper introduces SUPERB, a comprehensive benchmark for evaluating speech foundation models. The goal is to assess the generalizability of these models across diverse speech processing tasks. The benchmark includes 15 tasks, covering content, speaker characteristics, prosody, semantics, and generation. A unified framework is proposed, using a frozen foundation model with task-specific, lightweight prediction heads. The framework employs a learnable weighted-sum approach to combine representations from all layers of the foundation model, enabling effective performance across multiple tasks. The benchmark emphasizes the direct usability of SSL models on real-world applications, providing a standardized task design and evaluation protocol. The results show that SSL techniques are promising for building speech foundation models, with the best-performing models achieving competitive generalizability across most tasks. The benchmark also includes an online leaderboard for submissions and a community-driven database to support new development cycles. The paper analyzes the effectiveness of the framework, including information flows across tasks, the correctness of the weighted-sum benchmarking protocol, and the statistical significance and robustness of the benchmark. It also explores the performance of SSL models on different tasks, highlighting the importance of layer-wise analysis and the potential for improving performance through layer-wise benchmarking. The results demonstrate that SSL models can achieve strong task generalizability, with some models outperforming traditional non-SSL approaches. The benchmark provides a standardized evaluation framework for speech foundation models, enabling researchers to compare and improve their models effectively.